26/07/2019

Tay

Tay, the white supremacist

Meet Steve

Other uses of machine learning in society

  • HART (Harm Assessment Risk Tool)
  • COMPAS (Correctional Offender Management Profiling for Alternative Sanctions)
  • IBM Green Horizon
  • Tesla self driving cars
  • Virtual assisants
  • AlphaGo

Sample bias

  • collected data doesn’t accurately represent the environment the program is expected to run into.
    • self selection
    • undercoverage
    • survivorship
    • response

Sample bias - overcoming

  • stratification
  • weighting

Exclusion bias

  • result of excluding some features from our dataset usually under the umbrella of cleaning our data

Exclusion bias - overcoming

  • thoroughly investigating the features/groups we’re thinking of excluding beforehand
  • having colleagues cast an eye over variables
  • using the packages such as caret to calculate the relative feature importance

Observer bias

  • tendency to see what we expect to see, or what we want to see

Observer bias - overcoming

  • randomized controlled trials
  • double blind trials
  • train people as observers who have little or no stake in the outcome of the experiment

Prejudice bias

  • result of cultural influences or stereotypes.

Prejudice bias - overcoming

  • culmination of previous techniques to overcome
    • double blind trials
    • stratification
    • having people who aren’t the experiment designer do the sampling and others do the testing

Measurement bias

  • an issue with the device used to observe or measure.

Measurement bias - overcoming

  • multiple measuring devices
  • having trained people using the instruments

Identifying bias in training datasets

  • missing values
  • unexpected values
  • data skew

How to compensate for bias when building models

  • Lime
  • FairML
  • Google What/If…
  • IBM Bias assessment toolkit

How to monitor model for negative feedback

  • sister model for anomaly detection